Simple repeating patterns shift stereo depth by up to 20 meters.

Self-driving cars, drones, and robots rely on pairs of cameras, called stereo cameras, to judge how far away objects are. However, researchers discovered a built-in weakness in how these cameras sample pixels and calibrate their views. An attacker can exploit this weakness using simple repeating patterns, without needing advanced machine learning tricks. These patterns trick the software into miscalculating an object's true distance. The researchers showed this flaw affects classic matching algorithms, modern deep learning models, and commercial cameras. For instance, testing on a commercial camera showed an attacker could make an obstacle seem up to 20 meters farther away or 12 meters closer. In real-world driving tests, displaying this pattern for just 0.5 seconds triggered emergency braking in a popular self-driving system. Tests in a simulator also proved the attack works at speeds up to 40 km/h, tricking the systems into seeing false depths.

Why it matters

Self-driving cars and drones rely on stereo cameras for depth perception, but this research shows that simple repeating patterns can trick them into seeing obstacles much closer or farther than they are. A brief half-second attack can trigger emergency braking in a driving system, creating real collision risks. This vulnerability affects both traditional algorithms and modern deep learning models, raising serious safety concerns for autonomous systems.

Questions answered